Software Group © 2005 IBM Corporation Compiler Technology October 17, 2005 Array privatization in IBM static compilers -- technical report CASCON 2005

Compiler technology © 2005 IBM Corporation October 17, 2005 Guansong Zhang, Erik Charlebois and Roch Archambault Compiler Development Team IBM Toronto Lab, Canada Authors

Compiler technology © 2005 IBM Corporation October 17, 2005 Overview  Introduction and motivation  Array data flow analysis  Array data privatization  Performance results  Future work  Possible usage

Compiler technology © 2005 IBM Corporation October 17, 2005 Expose limitations  Compare SPEC2000FP and SPECOMP  SPECOMP achieves good performance and scalability –Compare between explicit and auto-parallelization  Expose missed opportunities  10 common benchmarks –Compare on a loop-to-loop basis

Compiler technology © 2005 IBM Corporation October 17, 2005 Improved auto-parallelization performance

Compiler technology © 2005 IBM Corporation October 17, 2005 Array privatization example

Compiler technology © 2005 IBM Corporation October 17, 2005 Basic loop parallelizer

Compiler technology © 2005 IBM Corporation October 17, 2005 Pre-parallelization Phase  Induction variable identification  Scalar Privatization --- only scalar !  Reduction finding  Loop transformations favoring parallelism

Compiler technology © 2005 IBM Corporation October 17, 2005 The concept of data privatization  Data is local to each loop iteration Do I = 1, 10 Temp = = … Temp … Enddo  Purpose: eliminating loop carried dependences.

Compiler technology © 2005 IBM Corporation October 17, 2005 The concept of data privatization (cont.)  Array as temp data do J = 1, 10 do I = 1, 10 Temp(I) =... end do do I = 1, = … Temp (I) … enddo

Compiler technology © 2005 IBM Corporation October 17, 2005 Array data flow and its structure  Similar to data flow –MayDef: array elements that may be written. –MustDef: array elements that are definitely written. –UpExpUse: array elements that may have an upward exposed use a use not preceded by a definition along a path from the loop header –LiveOnExit: array elements that are used after the loop region.  GARs: Guarded Array Regions (GARs). –A GAR is a tuple(G,D), D is a bounded Regular Section Descriptor (RSD) for the accessed array section, G is a guard that specifies the condition under which D is accessed  Notes: many papers discussed the issue

Compiler technology © 2005 IBM Corporation October 17, 2005 Array privatization algorithm

Compiler technology © 2005 IBM Corporation October 17, 2005 Loop normalization and array data flow  Normalized loop

Compiler technology © 2005 IBM Corporation October 17, 2005 Alias analysis and array data flow  Ideal situation: no alias at all. –Other wise, you can not tell what is the precise intersection of the two array section involved  Alias coming from: –Structural members, e.g. scalar replacement –Function parameters, array is a shadow (not mapped data, alias to any global array)  Procedure summary may help –Alias as fall back

Compiler technology © 2005 IBM Corporation October 17, 2005 Possible parallelization results

Compiler technology © 2005 IBM Corporation October 17, 2005 Real case

Compiler technology © 2005 IBM Corporation October 17, 2005 NAS MG (-O3 –qhot –q64)

Compiler technology © 2005 IBM Corporation October 17, 2005 Summary  Challenges – Compilation time Work with other optimizations – Loop unroll Graph complexity – Number of branches Array section caculation accuracy – Memory usage Managing and reusing

Compiler technology © 2005 IBM Corporation October 17, 2005 Summary (cont.)  Further improvement: – Inter-procedural array data flow Procedure summary More accurate section information instead of using alias – Symbolic range analysis Expression simplifier: lot of room to be improved – Compilation efficiency  Possible usage – Auto parallelization – Array contraction – Array coalescing